1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Generic sched_clock() support, to extend low level hardware time
4 * counters to full 64-bit ns values.
5 */
6#include <linux/clocksource.h>
7#include <linux/init.h>
8#include <linux/jiffies.h>
9#include <linux/ktime.h>
10#include <linux/kernel.h>
11#include <linux/math.h>
12#include <linux/moduleparam.h>
13#include <linux/sched.h>
14#include <linux/sched/clock.h>
15#include <linux/syscore_ops.h>
16#include <linux/hrtimer.h>
17#include <linux/sched_clock.h>
18#include <linux/seqlock.h>
19#include <linux/bitops.h>
20
21#include "timekeeping.h"
22
23/**
24 * struct clock_data - all data needed for sched_clock() (including
25 * registration of a new clock source)
26 *
27 * @seq: Sequence counter for protecting updates. The lowest
28 * bit is the index for @read_data.
29 * @read_data: Data required to read from sched_clock.
30 * @wrap_kt: Duration for which clock can run before wrapping.
31 * @rate: Tick rate of the registered clock.
32 * @actual_read_sched_clock: Registered hardware level clock read function.
33 *
34 * The ordering of this structure has been chosen to optimize cache
35 * performance. In particular 'seq' and 'read_data[0]' (combined) should fit
36 * into a single 64-byte cache line.
37 */
38struct clock_data {
39 seqcount_latch_t seq;
40 struct clock_read_data read_data[2];
41 ktime_t wrap_kt;
42 unsigned long rate;
43
44 u64 (*actual_read_sched_clock)(void);
45};
46
47static struct hrtimer sched_clock_timer;
48static int irqtime = -1;
49
50core_param(irqtime, irqtime, int, 0400);
51
52static u64 notrace jiffy_sched_clock_read(void)
53{
54 /*
55 * We don't need to use get_jiffies_64 on 32-bit arches here
56 * because we register with BITS_PER_LONG
57 */
58 return (u64)(jiffies - INITIAL_JIFFIES);
59}
60
61static struct clock_data cd ____cacheline_aligned = {
62 .read_data[0] = { .mult = NSEC_PER_SEC / HZ,
63 .read_sched_clock = jiffy_sched_clock_read, },
64 .actual_read_sched_clock = jiffy_sched_clock_read,
65};
66
67static __always_inline u64 cyc_to_ns(u64 cyc, u32 mult, u32 shift)
68{
69 return (cyc * mult) >> shift;
70}
71
72notrace struct clock_read_data *sched_clock_read_begin(unsigned int *seq)
73{
74 *seq = read_seqcount_latch(s: &cd.seq);
75 return cd.read_data + (*seq & 1);
76}
77
78notrace int sched_clock_read_retry(unsigned int seq)
79{
80 return read_seqcount_latch_retry(s: &cd.seq, start: seq);
81}
82
83static __always_inline unsigned long long __sched_clock(void)
84{
85 struct clock_read_data *rd;
86 unsigned int seq;
87 u64 cyc, res;
88
89 do {
90 seq = raw_read_seqcount_latch(s: &cd.seq);
91 rd = cd.read_data + (seq & 1);
92
93 cyc = (rd->read_sched_clock() - rd->epoch_cyc) &
94 rd->sched_clock_mask;
95 res = rd->epoch_ns + cyc_to_ns(cyc, mult: rd->mult, shift: rd->shift);
96 } while (raw_read_seqcount_latch_retry(s: &cd.seq, start: seq));
97
98 return res;
99}
100
101unsigned long long noinstr sched_clock_noinstr(void)
102{
103 return __sched_clock();
104}
105
106unsigned long long notrace sched_clock(void)
107{
108 unsigned long long ns;
109 preempt_disable_notrace();
110 /*
111 * All of __sched_clock() is a seqcount_latch reader critical section,
112 * but relies on the raw helpers which are uninstrumented. For KCSAN,
113 * mark all accesses in __sched_clock() as atomic.
114 */
115 kcsan_nestable_atomic_begin();
116 ns = __sched_clock();
117 kcsan_nestable_atomic_end();
118 preempt_enable_notrace();
119 return ns;
120}
121
122/*
123 * Updating the data required to read the clock.
124 *
125 * sched_clock() will never observe mis-matched data even if called from
126 * an NMI. We do this by maintaining an odd/even copy of the data and
127 * steering sched_clock() to one or the other using a sequence counter.
128 * In order to preserve the data cache profile of sched_clock() as much
129 * as possible the system reverts back to the even copy when the update
130 * completes; the odd copy is used *only* during an update.
131 */
132static void update_clock_read_data(struct clock_read_data *rd)
133{
134 /* steer readers towards the odd copy */
135 write_seqcount_latch_begin(s: &cd.seq);
136
137 /* now its safe for us to update the normal (even) copy */
138 cd.read_data[0] = *rd;
139
140 /* switch readers back to the even copy */
141 write_seqcount_latch(s: &cd.seq);
142
143 /* update the backup (odd) copy with the new data */
144 cd.read_data[1] = *rd;
145
146 write_seqcount_latch_end(s: &cd.seq);
147}
148
149/*
150 * Atomically update the sched_clock() epoch.
151 */
152static void update_sched_clock(void)
153{
154 u64 cyc;
155 u64 ns;
156 struct clock_read_data rd;
157
158 rd = cd.read_data[0];
159
160 cyc = cd.actual_read_sched_clock();
161 ns = rd.epoch_ns + cyc_to_ns(cyc: (cyc - rd.epoch_cyc) & rd.sched_clock_mask, mult: rd.mult, shift: rd.shift);
162
163 rd.epoch_ns = ns;
164 rd.epoch_cyc = cyc;
165
166 update_clock_read_data(rd: &rd);
167}
168
169static enum hrtimer_restart sched_clock_poll(struct hrtimer *hrt)
170{
171 update_sched_clock();
172 hrtimer_forward_now(timer: hrt, interval: cd.wrap_kt);
173
174 return HRTIMER_RESTART;
175}
176
177void sched_clock_register(u64 (*read)(void), int bits, unsigned long rate)
178{
179 u64 res, wrap, new_mask, new_epoch, cyc, ns;
180 u32 new_mult, new_shift;
181 unsigned long r, flags;
182 char r_unit;
183 struct clock_read_data rd;
184
185 if (cd.rate > rate)
186 return;
187
188 /* Cannot register a sched_clock with interrupts on */
189 local_irq_save(flags);
190
191 /* Calculate the mult/shift to convert counter ticks to ns. */
192 clocks_calc_mult_shift(mult: &new_mult, shift: &new_shift, from: rate, NSEC_PER_SEC, minsec: 3600);
193
194 new_mask = CLOCKSOURCE_MASK(bits);
195 cd.rate = rate;
196
197 /* Calculate how many nanosecs until we risk wrapping */
198 wrap = clocks_calc_max_nsecs(mult: new_mult, shift: new_shift, maxadj: 0, mask: new_mask, NULL);
199 cd.wrap_kt = ns_to_ktime(ns: wrap);
200
201 rd = cd.read_data[0];
202
203 /* Update epoch for new counter and update 'epoch_ns' from old counter*/
204 new_epoch = read();
205 cyc = cd.actual_read_sched_clock();
206 ns = rd.epoch_ns + cyc_to_ns(cyc: (cyc - rd.epoch_cyc) & rd.sched_clock_mask, mult: rd.mult, shift: rd.shift);
207 cd.actual_read_sched_clock = read;
208
209 rd.read_sched_clock = read;
210 rd.sched_clock_mask = new_mask;
211 rd.mult = new_mult;
212 rd.shift = new_shift;
213 rd.epoch_cyc = new_epoch;
214 rd.epoch_ns = ns;
215
216 update_clock_read_data(rd: &rd);
217
218 if (sched_clock_timer.function != NULL) {
219 /* update timeout for clock wrap */
220 hrtimer_start(timer: &sched_clock_timer, tim: cd.wrap_kt,
221 mode: HRTIMER_MODE_REL_HARD);
222 }
223
224 r = rate;
225 if (r >= 4000000) {
226 r = DIV_ROUND_CLOSEST(r, 1000000);
227 r_unit = 'M';
228 } else if (r >= 4000) {
229 r = DIV_ROUND_CLOSEST(r, 1000);
230 r_unit = 'k';
231 } else {
232 r_unit = ' ';
233 }
234
235 /* Calculate the ns resolution of this counter */
236 res = cyc_to_ns(cyc: 1ULL, mult: new_mult, shift: new_shift);
237
238 pr_info("sched_clock: %u bits at %lu%cHz, resolution %lluns, wraps every %lluns\n",
239 bits, r, r_unit, res, wrap);
240
241 /* Enable IRQ time accounting if we have a fast enough sched_clock() */
242 if (irqtime > 0 || (irqtime == -1 && rate >= 1000000))
243 enable_sched_clock_irqtime();
244
245 local_irq_restore(flags);
246
247 pr_debug("Registered %pS as sched_clock source\n", read);
248}
249EXPORT_SYMBOL_GPL(sched_clock_register);
250
251void __init generic_sched_clock_init(void)
252{
253 /*
254 * If no sched_clock() function has been provided at that point,
255 * make it the final one.
256 */
257 if (cd.actual_read_sched_clock == jiffy_sched_clock_read)
258 sched_clock_register(jiffy_sched_clock_read, BITS_PER_LONG, HZ);
259
260 update_sched_clock();
261
262 /*
263 * Start the timer to keep sched_clock() properly updated and
264 * sets the initial epoch.
265 */
266 hrtimer_setup(timer: &sched_clock_timer, function: sched_clock_poll, CLOCK_MONOTONIC, mode: HRTIMER_MODE_REL_HARD);
267 hrtimer_start(timer: &sched_clock_timer, tim: cd.wrap_kt, mode: HRTIMER_MODE_REL_HARD);
268}
269
270/*
271 * Clock read function for use when the clock is suspended.
272 *
273 * This function makes it appear to sched_clock() as if the clock
274 * stopped counting at its last update.
275 *
276 * This function must only be called from the critical
277 * section in sched_clock(). It relies on the read_seqcount_retry()
278 * at the end of the critical section to be sure we observe the
279 * correct copy of 'epoch_cyc'.
280 */
281static u64 notrace suspended_sched_clock_read(void)
282{
283 unsigned int seq = read_seqcount_latch(s: &cd.seq);
284
285 return cd.read_data[seq & 1].epoch_cyc;
286}
287
288int sched_clock_suspend(void)
289{
290 struct clock_read_data *rd = &cd.read_data[0];
291
292 update_sched_clock();
293 hrtimer_cancel(timer: &sched_clock_timer);
294 rd->read_sched_clock = suspended_sched_clock_read;
295
296 return 0;
297}
298
299static int sched_clock_syscore_suspend(void *data)
300{
301 return sched_clock_suspend();
302}
303
304void sched_clock_resume(void)
305{
306 struct clock_read_data *rd = &cd.read_data[0];
307
308 rd->epoch_cyc = cd.actual_read_sched_clock();
309 hrtimer_start(timer: &sched_clock_timer, tim: cd.wrap_kt, mode: HRTIMER_MODE_REL_HARD);
310 rd->read_sched_clock = cd.actual_read_sched_clock;
311}
312
313static void sched_clock_syscore_resume(void *data)
314{
315 sched_clock_resume();
316}
317
318static const struct syscore_ops sched_clock_syscore_ops = {
319 .suspend = sched_clock_syscore_suspend,
320 .resume = sched_clock_syscore_resume,
321};
322
323static struct syscore sched_clock_syscore = {
324 .ops = &sched_clock_syscore_ops,
325};
326
327static int __init sched_clock_syscore_init(void)
328{
329 register_syscore(syscore: &sched_clock_syscore);
330
331 return 0;
332}
333device_initcall(sched_clock_syscore_init);
334

source code of linux/kernel/time/sched_clock.c